沐阳 H3 · Turbo LoRA 联合音画加载调度

MiniMax H3 is a fast-approaching model, and Turbo LoRAs are how you stop it chewing through hours of sampling. But H3 isn't one sigma schedule - video and audio take different shift values, and getting the pair wrong silently corrupts your output. H3TurboSchedule exists to make that pairing painless: it merges ComfyUI's "LoRA loader (model only)" with the official H3 Sigma Shift, so one node loads the LightX2V Turbo LoRA and validates the whole parameter contract at once.

The mechanism is preset-driven. Pick a profile and the node knows the exact video/audio shift pair, step count, and the constraints that go with it:

| Profile | Video/audio shift | Steps | |---|---|---| | LightX2V v1.0 8-step | 12 / 3 | 8 (start here; 4 if you dare) | | LightX2V v1.0 4-step 768P | 6 / 3 | 4 | | LightX2V Ref2VA v0.1 4-step | 12 / 3 | 4 | | LightX2V v0.1 4-step | 12 / 3 | 4 |

The auto profile (recommended) reads the official ComfyUI LoRA filename and recognizes FL2VA vs Ref2VA, 4 vs 8 step, and 768P on its own. Prefer an FL2VA/T2VA preset for straight generation and a Ref2VA preset for motion transfer. There's also a 手动 manual mode where shift_video, shift_audio and recommended_steps actually take effect - that's for people who've independently validated their own shifts, not for beginners.

The important inputs:

• model - the base MODEL. Either wire a fresh model and load the LoRA here, or keep LoRA文件 on its default "not loading here (legacy workflow)" and feed in a model that an upstream LoraLoaderModelOnly already handled. Avoid double-loading.
• LoRA强度 - the equivalent of strength_model; LightX2V's examples sit at 1.0.
• speed_cache - optional TE-Speed (timestep cache, ~40%) or Spectrum (spectral, ~35%) hooks. Both need their own separately installed sibling node packs; if one's missing, the node warns and continues with it off.

Outputs are model (the patched model), plus recommended_steps, shift_video and shift_audio - wire those numbers into your H3 sampler and sampler settings so the schedule matches the LoRA. Note the contract Turbo enforces: fixed NFE trajectory, simple scheduler, denoise=1.0, Euler sampler. If your output looks over-baked or washed out, check those before touching anything else.

The trap the author calls out specifically: in v0.1.0, the upstream 8-step v1.0 768P checkpoint has no dedicated profile - do not rely on filename auto-detection for it. Stick to a listed preset or wait for the pack update that adds it. Also don't run this against a graph where you're simultaneously using a different H3 continuity implementation - the pack explicitly warns to use only one.

Install is the shared pack route: ComfyUI Manager → ComfyUI-MiniMaxH3-Myang → restart (hard-refresh if nodes look stale), or git clone https://github.com/civilcoco/ComfyUI-MiniMaxH3-Myang.git into custom_nodes. No extra Python deps; the LoRA itself comes from the LightX2V model page, and it rides on top of the H3 weights - which, reminder, are big and territory-restricted under the H3 Community License.

Start at 8 steps, watch a two-segment render, then decide if 4 is worth the speed. Turbo is a speed hack, not a quality hack - it's brilliant when it works, and this node exists so it works more often.